fastdds_transport_viz is a Command Line Interface (CLI) tool designed to predict and measure the DDS transport used by ROS 2 topics. It determines the transport mechanism (UDPv4/v6, TCP, SHM, or zero-copy data-sharing) for each topic based on Fast DDS discovery data. The tool can measure real traffic statistics, including rate and latency, using the Fast DDS statistics module. It offers various output options, including a live view, JSON output, and a web viewer. The tool has been verified across different ROS 2 distributions and hardware setups.
NASA is offering an opportunity for U.S. researchers to propose experiments using a 7-DoF robotic arm in low Earth orbit as part of the Space Roboticist Challenge. This involves submitting a white paper detailing a short, focused experiment to be conducted on the active system in microgravity. The process includes a simulation and validation phase with NASA engineers before offering on-orbit experiment time. Eligibility is restricted to U.S. citizens and permanent residents at eligible U.S. institutions with faculty sponsorship. Registration closes on September 23, 2026, with white paper deadlines in October 2026.
Beginners often struggle with workspace and sourcing issues when learning ROS 2 before writing any robot code. The author recorded the entire process, including errors, to address these common stumbling blocks. The video specifically covers workspace setup, package management, the first node, entry points, and the necessity of per-terminal sourcing. A key difficulty highlighted is understanding why sourcing is necessary for each terminal session. This content is intended for absolute first-week beginners to resolve initial setup problems.
Vision-Language-Action (VLA) policies need reliable failure detection for safe robotic execution, as existing methods suffer from label noise due to supervision mismatches. This work proposes a data-efficient framework to detect failures at a fine-grained timestamp level without dense annotation. The method first uses unlabeled action chunks to generate weak supervision signals capturing abnormal patterns like inconsistent actions. Active learning is then employed to select uncertain trajectories for annotation, fine-tuning the detector with these informative labels. Experiments demonstrate that the proposed method improves both timestamp-level and trajectory-level failure detection performance across various VLA policies.
Pretrained vision-language-action (VLA) models lack precision, which is addressed by VLA-Precision, an online reinforcement learning framework. It uses Asymmetric Co-Bootstrapping (ACoB) to efficiently improve VLA models through real-world experience. The ACoB-Stream architecture enables this process by decoupling states and streaming experience to handle large VLAs efficiently. Evaluations on chemistry tasks show VLA-Precision achieves a 98.3\% success rate in reduced time and increased sample efficiency. This method results in significant computational improvements over standard VLA and RL baselines.
Edge-assisted fusion and path prediction enhance planning for Connected Autonomous Vehicles (CAVs) by combining data from AVs and Road Side Units (RSUs). Conductor is presented as an edge-based solution for creating a unified world model and predicting AV trajectories while respecting a tight Age of Information (AoI) budget. The method uses an occlusion-aware selector to prioritize information from AVs detecting unseen objects. A runtime controller dynamically manages the number of AV inputs and predictions to ensure adherence to the time budget. Evaluations confirm this joint approach maintains safety bounds and high fusion fidelity in simulations with numerous CAVs.
AquaBEV is a monocular underwater occupancy model that predicts Bird's Eye View (BEV) occupancy from a single RGB image. It utilizes paired 3D imaging sonar data for geometric supervision during training. The model maps visual features into a calibration-free polar representation before reconstructing the prediction in Cartesian BEV coordinates. A unified protocol was established for an underwater occupancy benchmark using both RGB and sonar data. AquaBEV achieved relative improvements over baselines in visible and observed IoU metrics.
Sampling-based motion planning often claims asymptotic near-optimality by assuming the almost sure sampling of $\delta$-similar trajectories. This proof relies on the unstated assumption that $\delta$-similar trajectory segments will always be kept after sampling, which is not generally true. A scenario called "crowding out" demonstrates that locally low-cost paths can prevent $\delta$-similar trajectories from being added to the planning tree. Despite this issue, the paper shows that asymptotic near-optimality guarantees can still be achieved without $\delta$-similarity guarantees when crowding out is properly accounted for. An example is provided showing where inductively sampling a $\delta$-similar solution trajectory becomes impossible.
SocioGesture is a real-time, adaptive system designed to perceive social gestures for human-robot interaction. It utilizes a compact confidence-aware body-hand skeleton representation and a dual-stream model to fuse body motion and hand articulation for low-latency recognition. The model is trained with occlusion-aware skeleton corruption to enhance robustness against missing or occluded body parts. It operates in real-time on edge devices and saves uncertain segments for future offline adaptation to expand gesture vocabulary. This work provides a practical method for robust and efficient social gesture perception in robotics.
Continual Field-Adaptive Models (CFAMs) are introduced to enable post-deployment physical AI by allowing systems to learn efficiently after initial training. CFAMs use a complementary architecture with a frozen slow-learning component and a fast-learning Capsule Field for autonomous, gradient-free, on-device updates. The slow component includes modules for Sensor, Reasoning, and Action to handle multimodal input, task decomposition, and skill execution. Evaluation across various embodiments shows CFAMs achieve similar performance to full prior-trained policies while improving action success in near-OOD cases. Consequently, CFAMs offer a bounded approach to physical intelligence through few-shot skill learning and autonomous field growth.
NavArena is an automated framework that converts fixed 3D Gaussian Splatting (3DGS) reconstructions into benchmarks for goal-oriented visual navigation. It integrates a frozen 3DGS model for rendering, an occupancy costmap for collision queries, and semantic goal candidates to support autonomous navigation evaluation. The system automatically generates over 22.2 million expert trajectories across more than 2,000 scenes. Spatial and semantic evaluations assess the navigation representations derived from this process. NavArena aims to enable scalable and reproducible navigation evaluation on large-scale 3DGS data.
Personalized autonomous packing requires robots to understand resident preferences beyond scene geometry. The research investigates triadic human-robot collaboration involving a resident, a correction mediator, and a robot to handle these preferences. Two mediation methods, human-expert and voice-agent, were compared across Protection, Compactness, and Grouping tasks using a Show-Correct-Generalize process. Voice-agent mediation yielded comparable results to human-expert mediation in two preference categories, despite receiving less detailed instructions. This suggests voice agents can reduce expert involvement in personalized packing tasks.